Qwen3.6-35B-A3B benchmark tops AWQ Q4

The interesting part here is that the serving stack matters as much as the weight format. Once MTP and other runtime optimizations enter the picture, a “heavier” precision setup can still beat a lower-bit quantized one.

• –Serial decode came out at 110 tok/s for FP8 + MTP versus 91.8 tok/s for AWQ Q4
• –At concurrency 4, FP8 + MTP cleared 400+ tok/s while Q4 landed at 248 tok/s
• –At concurrency 8, FP8 + MTP hit 484 tok/s versus 250 tok/s for Q4
• –p90 latency at concurrency 8 improved from about 5.9s to about 3.4s
• –The comparison is not perfectly apples-to-apples because the Q4 setup lacked EP and MTP, which likely explains a lot of the gap